[4T3-OS-6d] OS-6

[4T3-OS-6d-01] Experimental investigation of XAI design for visual task

〇Akihiro Maehigashi¹, Yosuke Fukuchi², Seiji Yamada^2,3 (1. Shizuoka University, 2. National Institute of Informatics, 3. The Graduate University for Advanced Studies)

Keywords:Explainable AI (XAI), Trust, Explanation presentation

This study conducted two experiments to investigate interface designs that optimize the amount of explanation for explainable AI (XAI) in visual tasks. In Experiment 1, we used a classic visual search task to examine the effect of the explanation type on trust in the XAI. The results showed that (1) displaying only the AI results increased trust for a low-complexity task, and (2) displaying AI attention heatmaps with high interpretability increased trust for a high-complexity task. In Experiment 2, we tested the effect of an explanation display designed based on the results of Experiment 1 on task performance using an applied visual discrimination task. The results showed that (3) displaying only the AI results improved performance compared to the results with multiple explanations for a low-complexity task, and (4) displaying the AI results with AI attention heatmaps that had high interpretability improved performance compared to the results with multiple other explanations for a high-complexity task. This study showed the importance of an XAI interface design that optimally presents explanations for users.